Re: sched: context tracking demolishes pipe-test

CCs net-wizards since staring at profile/annotation isn't helping, and I
know all too well how these things laugh perversely at bisection.

On Tue, 2013-07-02 at 06:03 +0200, Mike Galbraith wrote: 

On Mon, 2013-07-01 at 11:20 +0200, Mike Galbraith wrote: 

quoted

On Mon, 2013-07-01 at 11:12 +0200, Mike Galbraith wrote: 

quoted

On Mon, 2013-07-01 at 10:06 +0200, Peter Zijlstra wrote:

quoted

So aside from the context tracking stuff, there's still a regression
we might want to look at. That's still a ~10% drop against 2.6.32 for
TCP_RR and few percents for tbench.

Yeah, known, and some of it's ours.

(btw tbench has a ~5% phase-of-moon jitter, you can pretty much
disregard that one)

Hm.  Seems we don't own much of TCP_RR regression after all, somewhere
along the line while my silly-tester hat was moldering, we got some
cycles back.. in the light config case anyway.

With wakeup granularity set to zero, per pipe-test, scheduler is within
variance of .32, sometimes appearing a tad lighter, though usually a wee
bit heavier.  TCP_RR throughput delta does not correlate.

echo 0 > sched_wakeup_granularity_ns

pipe-test
2.6.32-regress    689.8 Khz            1.000
3.10.0-regress    682.5 Khz             .989

netperf TCP_RR
2.6.32-regress   117910.11 Trans/sec   1.000
3.10.0-regress    96955.12 Trans/sec    .822

It should be closer than this. 

     3.10.0-regress                                              2.6.32-regress
     3.85%  [kernel]        [k] tcp_ack                          4.04%  [kernel]        [k] tcp_sendmsg
     3.34%  [kernel]        [k] __schedule                       3.63%  [kernel]        [k] schedule
     2.93%  [kernel]        [k] tcp_sendmsg                      2.86%  [kernel]        [k] tcp_recvmsg
     2.54%  [kernel]        [k] tcp_rcv_established              2.83%  [kernel]        [k] tcp_ack
     2.26%  [kernel]        [k] tcp_transmit_skb                 2.19%  [kernel]        [k] system_call
     1.90%  [kernel]        [k] __netif_receive_skb_core         2.16%  [kernel]        [k] tcp_transmit_skb
     1.87%  [kernel]        [k] tcp_v4_rcv                       2.07%  libc-2.14.1.so  [.] __libc_recv
     1.84%  [kernel]        [k] tcp_write_xmit                   1.95%  [kernel]        [k] _spin_lock_bh
     1.70%  [kernel]        [k] __switch_to                      1.89%  libc-2.14.1.so  [.] __libc_send
     1.57%  [kernel]        [k] tcp_recvmsg                      1.77%  [kernel]        [k] tcp_rcv_established
     1.54%  [kernel]        [k] _raw_spin_lock_bh                1.70%  [kernel]        [k] netif_receive_skb
     1.52%  libc-2.14.1.so  [.] __libc_recv                      1.61%  [kernel]        [k] tcp_v4_rcv
     1.43%  [kernel]        [k] ip_rcv                           1.49%  [kernel]        [k] native_sched_clock
     1.35%  [kernel]        [k] local_bh_enable                  1.49%  [kernel]        [k] tcp_write_xmit
     1.33%  [kernel]        [k] _raw_spin_lock_irqsave           1.46%  [kernel]        [k] __switch_to
     1.26%  [kernel]        [k] ip_queue_xmit                    1.35%  [kernel]        [k] dev_queue_xmit
     1.16%  [kernel]        [k] __inet_lookup_established        1.29%  [kernel]        [k] __alloc_skb
     1.14%  [kernel]        [k] mod_timer                        1.27%  [kernel]        [k] skb_release_data
     1.13%  [kernel]        [k] process_backlog                  1.26%  netserver       [.] recv_tcp_rr
     1.13%  [kernel]        [k] read_tsc                         1.22%  [kernel]        [k] local_bh_enable
     1.13%  libc-2.14.1.so  [.] __libc_send                      1.18%  netperf         [.] send_tcp_rr
     1.12%  [kernel]        [k] system_call                      1.18%  [kernel]        [k] sched_clock_local
     1.07%  [kernel]        [k] tcp_event_data_recv              1.11%  [kernel]        [k] copy_user_generic_string
     1.04%  [kernel]        [k] ip_finish_output                 1.07%  [kernel]        [k] _spin_lock_irqsave

	-Mike

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/